Land or Water?

Open models, asked 16,200 times each on one CPU. Pick a model and compare its map with the real Earth.

Repository Made with showtime

Model Answer key

↔

How to read this

P(Land) = P("Land") / (P("Land") + P("Water")) at the first answer token. Above 50% counts as a Land answer.
One prompt per point, empty cache, llama.cpp on CPU, 8-bit weights.
Area-weighted accuracy weights each point by cos(latitude), like the Claude chart, so the crowded polar rows do not dominate. The always-Water floor is what you get by answering Water everywhere.
Answer key: Natural Earth 1:10m land. Scores are recomputed in the repo by eval/score.py.
Keyboard: ← → cycle models. Link to a run with #run-id.